HIT-SCIR CoNLL2019 Unified Transition-Parser

This repository accompanies the paper, "HIT-SCIR at MRP 2019: A Unified Pipeline for Meaning Representation Parsing via Efficient Training and Effective Encoding", providing codes to train models and pre/post-precessing mrp dataset.

CoNLL2019 Shared Task Official Website: http://mrp.nlpl.eu/

Pre-requisites

Python 3.6
JAMR
NLTK
Gensim
Penman
AllenNLP 0.9.0

For JAMR installation, please refer to #2.

Dataset

Total training data is available at mrp-data.

Model

Download model from google-drive (CoNLL2019 Submission Version).

For prediction, please specify the BERT path in config.json to import the bert-indexer and bert-embedder. More prediction commands could be found in bash/predict.sh.

About BERT version, DM/PSD/UCCA/EDS use cased_L-12_H-768_A-12 (cased-bert-base) and AMR uses wwm_cased_L-24_H-1024_A-16 (wwm-cased-bert-large).

Usage

Prepare data

Step 1: Add companion to raw data.

We use conllu format companion data. This command adds companion.conllu to data.mrp and outputs to data.aug.mrp

python3 toolkit/augment_data.py \
    companion.conllu \
    data.mrp \
    data.aug.mrp

For evaluation data, you need to convert udpipe to conllu format and split raw input to 5 files. Run this command instead.

python3 toolkit/preprocess_eval.py \
    udpipe.mrp \
    input.mrp \
    --outdir /path/to/output

Step 2 (only for AMR): Convert data to amr format and run TAMR aligner.

Different from the other 4 parsers, our AMR parser accepts input of augmented amr format instead of mrp format.

Since TAMR's alignment is built on the JAMR alignment results, you need to set JAMR and CDEC path in bash/amr_preprocess.sh and run the command below.

bash bash/amr_preprocess.sh \
    data.aug.mrp \
    /path/to/word2wec

The final output is data.aug.mrp.actions.aug.txt which can be input to AMR parser.

According to TAMR, it is recommended to use the glove.840B.300d and filter the embeddings by the words and concepts (trimming the tail in word sense) in the data.

Train the parser

Based on AllenNLP, the training command is like

CUDA_VISIBLE_DEVICES=${gpu_id} \
TRAIN_PATH=${train_set} \
DEV_PATH=${dev_set} \
BERT_PATH=${bert_path} \
WORD_DIM=${bert_output_dim} \
LOWER_CASE=${whether_bert_is_uncased} \
BATCH_SIZE=${batch_size} \
    allennlp train \
        -s ${model_save_path} \
        --include-package utils \
        --include-package modules \
        --file-friendly-logging \
        ${config_file}

Refer to bash/train.sh for more and detailed examples.

Predict with the parser

The predicting command is like

CUDA_VISIBLE_DEVICES=${gpu_id} \
    allennlp predict \
        --cuda-device 0 \
        --output-file ${output_path} \
        --predictor ${predictor_class} \
        --include-package utils \
        --include-package modules \
        --batch-size ${batch_size} \
        --silent \
        ${model_save_path} \
        ${test_set}

More examples in bash/predict.sh.

Package structure

bash/ command pipelines and examples
config/ Jsonnet config files
metrics/ metrics used in training and evaluation
modules/ implementations of modules
toolkit/ external libraries and dataset tools
utils/ code for input/output and pre/post-processing

Acknowledgement

Thanks to the task organizers and also thanks to the developer of AllenNLP, JAMR and TAMR.

Contacts

For further information, please contact lxdou@ir.hit.edu.cn, yxu@ir.hit.edu.cn

longxudou/HIT-SCIR-CoNLL2019